Fixed-Point Implementation of Convolutional Neural Networks for Image Classification

In this paper, we show step-by-step how to design and optimize a fixed-point convolutional neural network (CNN) classifier. Moreover, the fixed-point classifier has been implemented using C++ programming and onto an FPGA. We show that the classifier with 4-bit fixed-point arithmetic with 8bit additions can classify handwritten digits with over 99.16% accuracy. Compared with the floating-point classifier which achieves 99.55% accuracy, the degradation due to fixed-point implementation is less than 0.4%.